A hybrid setup with multi-GPU servers (e.g., NVIDIA DGX) for training and edge devices (e.g., NVIDIA Jetson) for inference is most appropriate. Multi-GPU servers handle continuous training on large datasets with high compute power, while edge devices enable low-latency inference for real-time fraud detection, balancing scalability and speed. Option A (single GPU) lacks scalability. Option B (edge-only ARM) can't handle training demands. Option D (CPU-based) sacrifices GPU acceleration. NVIDIA's fraud detection architectures endorse this hybrid model.